lmb-freiburg / flownet2

FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
https://lmb.informatik.uni-freiburg.de/Publications/2017/IMKDB17/
Other
1k stars 318 forks source link

Questions about Kiiti training #199

Closed deepcharle closed 5 years ago

deepcharle commented 5 years ago

Thank you for your excellent work in optical flow. Note that the groundtruth flow values in Kiiti is very sparse, how do you deal with this case when you augment the image pairs and corresponding optical flow to keep it from interpolating with nonsense zero optical flow values?

nikolausmayer commented 5 years ago

If I remember correctly, we do not use geometric augmentation for KITTI. I'm pretty sure that our data augmentation cannot deal with NaN pixels.

The KITTI setting is so tightly constrained that any change to the data bears a significant risk of destroying the advantage you hope to gain by finetuning. Stretching, warping, or rotating images makes no sense because the KITTI test set has none of these effects.

deepcharle commented 5 years ago

thank you very much for your reply